The 3 Echidna Senior Staff completed the v1 diagnostic between July 22 and 27. Here's what I see in your responses, what you already agree on, and the decisions it sets us up to start mulling on. The full team picture will color in after Accra retreat.
You're skeptical and open at the same time, and the skepticism is specific to these areas: data security, accuracy, and what heavy AI use might do to human thought. No score in the group came in above 3 out of 5, which is telling. You see the largest efficiency gains in drafting, research scans, formatting, and admin, and you want tight guardrails before anything goes all-staff. Governance came through as non-negotiable in all three responses, which makes sense.
This was the dominant concern. What happens as AI tools learn Echidna's operations, and who else could use that knowledge?
"I worry as it learns about the operations of Echidna it could know quite a lot and could be leveraged to commit fraud or ransom the organization or risk the safety of our staff as they travel abroad."
2/3 raised a similar concern: the fear that AI becomes a shortcut for the engagement that Echidna's work depends on, like getting report summaries instead of reading what grantees submitted.
"I'm really worried about people using AI as a substitute for their own learning and engagement, which would be a detriment to our goals."
Trust in output is low, from senior staff's own experiences of hallucinations. For example, once Gemini pulled from three-year-old emails and drew dated conclusions. Another example saw terrible output on executive summaries and spreadsheet analysis.
"It takes real intentionality and skill to use it well and to be able to scrutinize what it spits out, discern whether it is good and accurate, and revise as needed. Not everyone applies that intentionality equally."
You each rated yourselves 1 (low) to 5 (high) on six capabilities. Shown here as the three individual scores, in the same order for every row.
| CAPABILITY | SCORES | WHAT I TAKE FROM IT |
|---|---|---|
| Technical understanding of AI models | 2 · 1 · 2 | Nobody above 2. Training starts from fundamentals, no jargon. |
| What AI can and can't do reliably | 2 · 3 · 2 | Low-moderate. Pairs with the accuracy theme above. |
| Writing prompts that get useful results | 2 · 3 · 2 | Room to grow quickly. This is very teachable. |
| Critically evaluating AI claims | 2 · 4 · 4 | A real split. One of you wants this skill; two already have it. |
| Knowing what's safe to share with AI | 1 · 4 · 4 | The widest split in the set. The policy work closes this gap for everyone. |
| Clarity on which tasks AI could help with | 2 · 2 · 2 | Unanimous, and exactly what this engagement is for. |
2 things stood out: All of you rate your confidence with identifying "which of my tasks could AI help with" at 2, so the task-mapping work in training will be vital. And knowing what's "safe to share" split widely (1 against 4 and 4), which says the data rules need to be explicitly laid out; they need to be written down, taught and imbibed, and easy to check.
This finding will impact the policy, and it's an important one to call out, which your team has already noticed, and it also describes almost every organization I see at this stage.
The team's instincts are solid: sensitive work is already being steered toward the enterprise accounts. But right now the boundary between "safe account" and "public account" depends on each person's individual judgment and settings. Written account rules and the software audit we added at kickoff (Zoom, Grammarly, Leapsome, Expensify, Bill.com, etc) will close that gap.
I thought the four scenario questions produced the richest material in the diagnostic. Across all three of you, the same rules kept appearing throughout:
One response closed with five principles (and I'd love to review the Echidna principles as well) - this list is close to a draft preamble for the policy.
"• Only as a way to enhance judgment and learning, not in ways that circumvent it. • Never in replacement for human interaction with grantees, colleagues, etc. • Human review and judgment are required for every product produced by AI -- ultimate responsibility lies with a human, and AI cannot be blamed for shortcomings. • No sensitive information to be shared with public versions of AI and other necessary controls for safeguarding private information. • Used only where it adds value, not just because it is there (mindful of environmental costs); if there are ways we can account more intentionally for the full externality costs of AI use (environmental, etc.) that would be great"
Does this track for everyone in this team? If so, my plan is to then weave these into the September 1 policy session as the starting text, so we begin from what you've already agreed on.
Update, Aug 3: Emily shared the 2022 Echidna Principles, and they strengthen this section considerably. Your principles already commit Echidna to reading and reacting to the materials grantees share, in acknowledgment of the time and effort involved in keeping you informed. That settles the scenario 3 question at the principle level: reading grantee materials is part of how Echidna honors grantee effort, and the AI policy can cite it directly. The full mapping from your principles to the policy lives in the Sep 1 pre-read.
The range exists, but it's fairly narrow. One of you is most cautious (2 and 2, weekly use), one most open (3 and 3, daily use), and one in between. The caution in these numbers runs a bit deeper than what I clocked in the kickoff, which is important to note. It doesn't change the plan, though, it just gives me a deeper understanding of pace and that proof will matter more than theories.
Scenario 3 asked about a program lead using AI to help assess a grant application. Your three answers land in slightly different places, and the differences matter for the policy:
"That is fine, if thoroughly reviewed by the team member."
"I do not think it is an appropriate use of AI to circumvent staff members own personal review of grant materials. The way we are structured, by the time we invite a grant application, we are giving an organization very serious consideration and what they submit requires human review -- this is one important tool in how we get to know an organization. I would be ok if the staff member is using AI to supplement their own review -- if they fed as a prompt their own assessment and asked it to find gaps/holes in their own analysis, for example. I also think we could rely more on AI for reviewing safeguarding policies and financial reports for risks because there are more obvious ways to train AI on what to look for on these. But it should never be used as a substitute for actual engagement with the materials."
"I don't think I'm bothered by this assuming the program lead is inputting (voicing/typing) their thoughts/impressions on the strengths, risks, ideas, recommendations they identified in reading the application themselves and AI is helping organize their rough ideas into a rubric/format the individual or org has created. I'd also feel comfortable with them asking AI to supplement their thinking or suggest anything they missed. I would want to ensure the information is secured within our enterprise account and not available or feeding models in general."
All three of you require human review. The difference is where the review sits. One answer accepts AI doing the first read as long as the person thoroughly reviews the output afterward. The other two require the person to read the source materials first, with AI only allowed to organize or gap-check their own thinking afterward.
The question to settle together: for grant materials specifically, is a thorough review of AI's output enough, or must the human read first, every time? The policy will have to pick one, because "use your judgment" is often the kind of ambiguity that results in usage drift.
Environmental impact came up in two of the three responses. Consistent with the steer I got at kickoff, so we'll pay attention and account for it in tool guidance without trying to over-quantify.
The tool decision will get made on August 21, after the full team completes the diagnostic in Accra. While we're a ways away from a decision still, here's the direction the leadership evidence suggests, and the live question we're still weighing out:
The case for Gemini first is practical. You already pay for it, it sits inside the Workspace you run on, one of you already does sensitive work in it, and Workspace enterprise data protections cover it.
Claude can play a big role for complex judgment work like grant synthesis and long-form analysis, and two of you already prefer it personally. If it enters Echidna more formally, we'd want to do it deliberately, on an enterprise footing, once the policy is live.
Verified pricing as of this week:
The open question: one tool for everyone, or a governed menu?
You asked us to recommend a single tool, and that may still be the answer in the end, but I wanted to flag now that some strong AI-adopting organizations run this a bit differently: they allow a small set of approved tools for people to choose between, all inside one policy, one data standard, and one shared prompt library. It costs a little more on coordination but can return better fit per person or team. The staff diagnostic will tell us which shape fits Echidna best, and we'll bring a firm recommendation to the August 21 session.
We're setting the first red lines on July 29. Your own responses already draw several of them, so here's the list as I'd start it:
A few open starter threads from the diagnostic and kickoff that could become red lines, or could stay flexible:
Did anything come up for you before or after taking the diagnostic that belongs on this list? It's better we over-collect red line candidates today and trim at the policy session together.
The full timeline with booking links lives on the schedule page.
Based on three leadership diagnostic responses, July 22–27, 2026. Quotes are verbatim and shown without names. Analysis by SA Advisory (Shabnam + Sara). Please shout if anything looks off; this page updates as the staff wave lands.